In _calculate_responses_token_usage, manual output token counting only handles streaming_message_responses but lacks a fallback for non-streaming responses. The completions version (line 220-223) has elif hasattr(response, 'choices') to extract output from non-streaming responses, but the responses version has no equivalent elif hasattr(response, 'output') branch. When the Responses API doesn't provide usage data and streaming_message_responses is None (non-streaming case at line 637), output tokens will incorrectly remain 0 instead of being manually counted.

In _calculate_responses_token_usage (lines 291-295), when output_tokens == 0 and streaming_message_responses is None, there's no fallback to manually count output tokens from response.output. The analogous function _calculate_completions_token_usage has this fallback (lines 220-223) which counts from response.choices. For non-streaming Responses API calls where the API doesn't return usage data and tiktoken is configured, output tokens won't be recorded.